CN102819836B - Method and system for image segmentation - Google Patents
Method and system for image segmentation Download PDFInfo
- Publication number
- CN102819836B CN102819836B CN201210224358.3A CN201210224358A CN102819836B CN 102819836 B CN102819836 B CN 102819836B CN 201210224358 A CN201210224358 A CN 201210224358A CN 102819836 B CN102819836 B CN 102819836B
- Authority
- CN
- China
- Prior art keywords
- sub
- subregion
- segmentation result
- segmentation
- features
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Expired - Fee Related
Links
Landscapes
- Image Analysis (AREA)
Abstract
The embodiment of the invention provides a method and a system for image segmentation. The method comprises the following steps: pre-segmenting a given image to obtain a plurality of sub-areas; merging the plurality of sub-areas according to at least two features to obtain sub-segmentation result corresponding to each feature; indicating sub-segmentation results corresponding to at least two features as hypergraphs; and performing cluster ensemble on the sub-segmentation results of the at the least two features based on the hypergraphs to obtain the segmentation result for the given image. The method is effectively combined with segmentation for the image based on multiple features, overcomes the defect that a common segmentation algorithm has no generality based on single image feature and avoids the problem that the features have no due effect on segmentation performance due to improper setting of visual feature vector and partial weights and dimensions in Bag of words.
Description
Technical field
The present invention relates to technical field of image processing, particularly relate to a kind of image partition method and system.
Background technology
Iamge Segmentation is the first step of image procossing and analysis, is the core technology of images steganalysis, is also simultaneously one of problem the most ancient and the most difficult in image procossing.Iamge Segmentation is the set representing the image as physically significant connected region, namely according to the priori of object and background, target in image, background are marked, located, then by process that target is separated from background or other pseudo-targets.
The method of Iamge Segmentation has multiple, any partitioning algorithm based on single features all can only produce good segmentation effect in the image class of this feature-sensitive, the hydraulic performance decline then split for the unconformable image of this feature is very large, therefore, merge manifold partitioning algorithm and could all obtain ideal segmentation performance to most image.
In prior art, adopt and the multiple global characteristics extracted are formed the mode of the proper vector of a higher-dimension to merge the dividing method of multiple features, dimension shared by often kind of visual signature is different, therefore the feature that dimension is high often serves leading role in segmentation, and further feature is difficult to play a role, the performance of segmentation is difficult to improve; Adopt the mode based on word bag (Bag of words) that the multiple global characteristics extracted and local feature are formed the partitioning scheme that vision word merges multiple features, because global characteristics is different with local characteristic properties, the visual word formed also is visibly different, therefore, these two kinds of visual word directly joined together simply to use the effect that can not play often kind of feature in segmentation, the performance of segmentation is difficult to improve.Because arranging of each several part dimension in visual feature vector and Bag of words and weights is improper, make the performance of feature to segmentation cannot play due effect, therefore, these modes cannot play the effect that often kind of feature plays in segmentation.
Summary of the invention
The embodiment of the present invention provides a kind of image partition method and system, can the manifold segmentation result of more effective fusion.
In order to solve the problems of the technologies described above, the technical scheme of the embodiment of the present invention is as follows:
A kind of image partition method, comprising:
Pre-segmentation is carried out to Given Graph picture, obtains multiple subregion;
Respectively described multiple subregion is merged according at least two kinds of features, obtain the sub-segmentation result that often kind of feature is corresponding;
Sub-segmentation result corresponding for described at least two kinds of features is expressed as hypergraph;
Based on described hypergraph, the sub-segmentation result of described at least two kinds of features is carried out clustering ensemble, obtain the segmentation result of described Given Graph picture.
Further, described described multiple subregion to be merged, comprising:
Determine the link area of every sub regions;
Calculate the merging weights of the link area of described every sub regions;
The link area that described every sub regions and its merging weights satisfy condition is merged.
Further, the described link area determining every sub regions, comprising:
Extract the eigenwert that subregion is adjacent subregion;
According to the eigenwert extracted, calculate described subregion and be adjacent characteristic similarity between subregion;
Characteristic similarity is greater than similarity threshold, and region area is greater than the link area of adjacent subregion as described subregion of described subregion.
Further, the merging weights of the link area of the described every sub regions of described calculating, comprising:
According to the characteristic similarity between described every sub regions and its link area and semantic similarity valuation functions, according to webpage level algorithm, determine the merging weights of the link area of described every sub regions.
Further, described semantic similarity valuation functions is normal distyribution function or polygronal function.
Further, described the link area that described every sub regions and its merging weights satisfy condition to be merged, comprising:
The link area of described every sub regions and its merging maximum weight is merged.
Further, describedly based on described hypergraph, the sub-segmentation result of described at least two kinds of features is carried out clustering ensemble, obtains the segmentation result of described Given Graph picture, comprising:
Set final clusters number;
Based on described hypergraph, the sub-segmentation result of described at least two kinds of features is carried out spectral clustering according to described final clusters number integrated, obtain the segmentation result of described Given Graph picture.
Further, the final clusters number of described setting comprises:
The sub-segmentation result corresponding for described at least two kinds of features calculates characteristic similarity respectively;
Be normalized calculating the characteristic similarity obtained;
Using the clusters number of the sub-segmentation result corresponding to the minimum value of characteristic similarity after normalization as described final clusters number.
A kind of image segmentation system, comprising:
Pre-segmentation unit, for carrying out pre-segmentation to Given Graph picture, obtains multiple subregion;
Merge cells, for merging described multiple subregion respectively according at least two kinds of features, obtains the sub-segmentation result that often kind of feature is corresponding;
Converting unit, for being expressed as hypergraph by sub-segmentation result corresponding for described at least two kinds of features;
Cutting unit, for the sub-segmentation result of described at least two kinds of features being carried out clustering ensemble based on described hypergraph, obtains the segmentation result of described Given Graph picture.
Further, described merge cells comprises:
Locator unit, for determining the link area of every sub regions;
Computation subunit, for calculating the merging weights of the link area of described every sub regions;
Merge subelement, merge for the link area that described every sub regions and its merging weights are satisfied condition.
Further, described locator unit comprises:
Extraction module, is adjacent the eigenwert of subregion for extracting subregion;
First computing module, for according to the eigenwert extracted, calculates described subregion and is adjacent characteristic similarity between subregion;
Determination module, for characteristic similarity is greater than similarity threshold, and region area is greater than the link area of adjacent subregion as described subregion of described subregion.
Further, described computation subunit, specifically for according to the characteristic similarity between described every sub regions and its link area and semantic similarity valuation functions, according to webpage level algorithm, determines the merging weights of the link area of described every sub regions.
Further, described merging subelement, specifically for merging the link area of described every sub regions and its merging maximum weight.
Further, described cutting unit comprises:
Setting subelement, for setting final clusters number;
Segmentation subelement, integrated for the sub-segmentation result of described at least two kinds of features being carried out spectral clustering according to described final clusters number based on described hypergraph, obtain the segmentation result of described Given Graph picture.
Further, described setting subelement comprises:
Second computing module, for calculating the characteristic similarity of sub-segmentation result corresponding to described often kind of feature;
Processing module, the characteristic similarity for the sub-segmentation result to described at least two kinds of features is normalized;
Setting module, for using the clusters number of the sub-segmentation result corresponding to the minimum value of characteristic similarity after normalization as described final clusters number.
The sub-segmentation result of different characteristic, by first obtaining manifold sub-segmentation result, then represents in a hypergraph, and then utilizes the method for clustering ensemble to obtain final image segmentation result by the embodiment of the present invention.This dividing method combines the segmentation of various features to image effectively, overcome general partitioning algorithm because do not possess the shortcoming of versatility based on single characteristics of image, simultaneously, by a hypergraph model, the segmentation result under various features space is merged, formed one of image final division by the clustering ensemble on hypergraph, avoid and make the performance of feature to segmentation cannot play effective problem of answering because of each several part weights in visual feature vector and Bag of words and the arranging improper of dimension, to noise, abnormity point, sampled point variation is insensitive, there is stronger robustness.
Accompanying drawing explanation
In order to be illustrated more clearly in the embodiment of the present invention or technical scheme of the prior art, be briefly described to the accompanying drawing used required in embodiment or description of the prior art below, apparently, accompanying drawing in the following describes is only some embodiments of the present invention, for those of ordinary skill in the art, under the prerequisite not paying creative work, other accompanying drawing can also be obtained according to these accompanying drawings.
Fig. 1 is the method flow diagram of a kind of Iamge Segmentation of the embodiment of the present invention;
Fig. 2 is a kind of method flow diagram merged multiple subregion of the embodiment of the present invention;
Fig. 3 is a kind of method flow diagram determining the link area of a certain subregion of the embodiment of the present invention;
Fig. 4 is the schematic diagram in the embodiment of the present invention, the sub-segmentation result of three kinds of features being expressed as hypergraph;
To be that the embodiment of the present invention is a kind of carry out the integrated method flow diagram of spectral clustering based on hypergraph antithetical phrase segmentation result to Fig. 5;
Fig. 6 is a kind of method flow diagram setting final clusters number of the embodiment of the present invention;
Fig. 7 is the structural representation of a kind of image segmentation system of the embodiment of the present invention;
Fig. 8 is the structural representation of a kind of merge cells of the embodiment of the present invention;
Fig. 9 is the structural representation of a kind of cutting unit of the embodiment of the present invention.
Embodiment
In order to make those skilled in the art can understand feature of the present invention and technology contents further, refer to following detailed description for the present invention and accompanying drawing, accompanying drawing only provides reference and explanation, is not used for limiting the present invention.
Below in conjunction with drawings and Examples, technical scheme of the present invention is described.
See Fig. 1, it is the method flow diagram of a kind of Iamge Segmentation of the embodiment of the present invention.
The method can comprise:
Step 101, carries out pre-segmentation to Given Graph picture, obtains multiple subregion.
In embodiments of the present invention, first pre-segmentation is carried out to Given Graph picture, be partitioned into multiple subregion.The method of segmentation can see multiple prior art, such as, name is utilized to be called " Normalized cuts and image segmentation ", author is J.Shi, and J.Malik., be published in IEEE Transactions on Pattern Analysis and Machine Intelligence, the method for the list of references proposition of 22 (8) (2000) 888-905, namely Iamge Segmentation is become several latticed join domains by Ncut method.
Step 102, merges multiple subregion respectively according at least two kinds of features, obtains the sub-segmentation result that often kind of feature is corresponding.
First, choose at least two kinds of features, such as color, texture, SIFT(Scale-invariant feature transform, scale invariant feature is changed) etc., then respectively multiple subregions that pre-segmentation obtains are merged based on different features, obtain the sub-segmentation result that often kind of feature is corresponding, such as, merge multiple subregions that pre-segmentation obtains according to color characteristic, the result after merging is sub-segmentation result corresponding to this color characteristic.
Wherein, for the merging of subregion, webpage level algorithm can be used or other are similar to the mode of webpage level algorithm, as the random walk method based on similarity, merging subregion, can also be adopt to merge subregion based on the revised webpage level algorithm of semantic similarity.Specifically refer to the description of subsequent embodiment.
Step 103, is expressed as hypergraph by sub-segmentation result corresponding for these at least two kinds of features.
After obtaining sub-segmentation result corresponding to each feature, sub-segmentation result corresponding for all features can be expressed as a hypergraph.Wherein, all subregion that pre-segmentation obtains is expressed as the summit of hypergraph, sub-segmentation result corresponding for each feature is expressed as the super limit of hypergraph.
The concrete process being converted into hypergraph refers to the description of subsequent embodiment.
Step 104, carries out clustering ensemble based on this hypergraph by the sub-segmentation result of at least two kinds of features, obtains the segmentation result of Given Graph picture.
After the sub-segmentation result of each feature is expressed as hypergraph, clustering ensemble can be carried out according to this hypergraph, obtain the final segmentation result of this Given Graph picture.Wherein, the method that clustering ensemble adopts can be spectral clustering, and also can be other clustering method, such as, K-means clustering method, complicated drawing method etc., will not enumerate herein.
The sub-segmentation result of different characteristic, by first obtaining manifold sub-segmentation result, then represents in a hypergraph, and then utilizes the method for clustering ensemble to obtain final image segmentation result by the embodiment of the present invention.This dividing method combines the segmentation of various features to image effectively, overcome general partitioning algorithm because do not possess the shortcoming of versatility based on single characteristics of image, simultaneously, by a hypergraph model, the segmentation result under various features space is merged, formed one of image final division by the clustering ensemble on hypergraph, avoid and make the performance of feature to segmentation cannot play effective problem of answering because of each several part weights in visual feature vector and Bag of words and the arranging improper of dimension, to noise, abnormity point, sampled point variation is insensitive, there is stronger robustness.The embodiment of the present invention provides the very strong dividing method of a kind of versatility, for dissimilar image, stability and the versatility of splitting quality can be ensured, and, the method is a kind of unsupervised auto Segmentation technology, without the need to inputting a large amount of training sets in advance, also without the need to artificial participation in segmentation, as specified the quantity of cluster, the seed amount etc. that appointed area increases.
Respectively multiple subregion is merged according at least two kinds of features in execution step 102, when obtaining sub-segmentation result corresponding to often kind of feature, the step obtaining sub-segmentation result for its merging subregion of often kind of feature is identical, in another embodiment of the invention, example is characterized as with one, as color characteristic, its process that multiple subregion is merged, as shown in Figure 2, can comprise:
Step 201, determines the link area of every sub regions.
In the present embodiment, can determine the link area of every sub regions according to the characteristic similarity between subregion and area, to determine the link area of a certain subregion, its detailed process as shown in Figure 3, can comprise:
Step 301, extracts the eigenwert that subregion is adjacent subregion.
Feature selected is in the present embodiment color, then namely extract the color feature value that subregion is adjacent subregion in this step, this leaching process see Eigenvalue Extraction Method of the prior art, can repeat no more herein.Wherein, adjacent with a certain sub-window position subregion is the adjacent subregion of this certain subregion.
Step 302, calculating subregion is adjacent the characteristic similarity between subregion.
The characteristic similarity that subregion is adjacent between subregion can calculate according to the eigenwert of two subregions, concrete, can be the Euclidean distance of the eigenwert of two subregions, the such as Euclidean distance of the color feature value of two subregions.
Step 303, is greater than similarity threshold by characteristic similarity, and region area is greater than the link area of adjacent subregion as this subregion of this subregion.
When the subregion characteristic similarity be adjacent between subregion is greater than the similarity threshold of setting, think that this subregion is adjacent between subregion and has linking relationship, then compare the region area of two subregions, namely the adjacent subregion being greater than this subregion area regards as the link area of this subregion.The link direction of subregion points to its link area.
Wherein, similarity threshold can rule of thumb set, and also can set respectively for different features, and such as, the similarity threshold that color characteristic is corresponding can be set as the characteristic similarity average of the hsv color of all subregions; The similarity threshold that textural characteristics is corresponding can be set as the characteristic similarity average of the texture co-occurrence matrix of all subregions; Similarity threshold corresponding to SIFT feature can be set as the SIFT feature similarity average of all subregions.
The link area of every sub regions can be determined by above-mentioned steps 301 ~ 302, in other embodiments, other method also can be adopted to determine the link area of every sub regions, such as, select adjacent boundary sizes etc. to determine.
After determining link area, perform following steps 202.
Step 202, calculates the merging weights of the link area of every sub regions.
The segmentation of present image is main it is considered that the bottom visual signature of image, these features are an attribute of image, and not there is any semantic meaning, relation one to one is not possessed with in the semantic understanding aspect of people, namely semantic gap problem, namely bottom visual signature is similar, does not represent semantic content consistent; And the object of semantic congruence, region also may on bottom visual signature difference very large.
In order to solve the semantic gap problem between vision low-level image feature and high-level semantic, this embodiment introduces semantic similarity and carry out alternative features similarity, namely have employed semantic similarity valuation functions, characteristic similarity is inputted semantic similarity valuation functions, this function simulates the relation of characteristic similarity and semantic similarity effectively, i.e. feature similarity not necessarily semantic similitude, and semantic similitude also may be characteristically different.
After introducing semantic similarity, the present embodiment have employed the merging weights being similar to webpage rank (Pagerank) the algorithm determination link area that web page joint uses, and then feasible region merges, the method not only considers the syntople in region, especially the merging of neighboring region is converted to the merging process based on semanteme, greatly improves the performance of Iamge Segmentation.
Concrete, according to the characteristic similarity between every sub regions and its link area and semantic similarity valuation functions, according to webpage level algorithm, the merging weights of the link area of every sub regions can be determined.
First, based on the Pagerank algorithm being similar to web page joint use, if subregion p
isubregion p has been pointed in link
j, be also subregion p
jfor subregion p
ilink area, then subregion p
idistribute to subregion p
jsemantic similarity be: subregion p
iand p
jsemantic similarity account for p
iwith the ratio of the semantic similarity sum in its all-links region, also namely:
Wherein, σ
s(p
i, p
j) be subregion p
iand p
jsemantic similarity, N (p
i) be a function, namely ask subregion p
iall-links region, subregion p
ieach link area p
krepresent.
The semantic similarity that the present embodiment utilizes the visual signature similarity between subregion (as color, Texture eigenvalue similarity) to come between approximate subregion, that is:
Wherein, sim (p
i, p
j) be p
i, p
jthe visual signature similarity (as color characteristic similarity, textural characteristics similarity, SIFT feature similarity etc.) of two link areas; E is a semantic similarity valuation functions, and E can select normal distyribution function or polygronal function, and other valuation functions certainly also may be selected to substitute.
Then, the subregion of image is merged and can regard the process being jumped to another node by a node according to the semantic similarity probability of hinged node as, then adopt the redirect probability be similar in Pagerank algorithm, the merging weights P of the link area of every sub regions can be obtained
r, that is:
Wherein, P
r(p
j) represent region P
jmerging weights, ε is a regulating constant, and n is the number of the subregion obtained after carrying out pre-segmentation to image, such as experiment in be set as 50, N (p
j) be a function, namely ask subregion p
jall-links region, subregion p
jeach link area p
krepresent.
The merging weights P of link area of every sub regions is being obtained based on above-mentioned formula
rafter, perform following steps 203.
Step 203, merges the link area that every sub regions and its merging weights satisfy condition.
According on step method calculate obtain every sub regions link area merging weights after, have employed in the present embodiment greed strategy, the link area by subregion and its merging maximum weight merges, and therefore accelerates the speed of merging.In other embodiments, these merging weights meet condition can as required or experience setting.
After execution above-mentioned steps 201 ~ 203, the sub-segmentation result that color characteristic is corresponding can be obtained, in like manner, sub-segmentation result corresponding to textural characteristics can be obtained, the sub-segmentation result that SIFT feature is corresponding according to above-mentioned steps.
After obtaining sub-segmentation result corresponding to some features, this little segmentation result can be expressed as hypergraph, all subregion that pre-segmentation obtains be expressed as the summit of hypergraph, sub-segmentation result corresponding for often kind of feature is expressed as the super limit of hypergraph.The method of this structure hypergraph similarly to the prior art, can be called see name " Cluster ensembles-a knowledge reuse framework for combining multiple partitions ", author is A.Strehl and J.Ghosh., be published in The Journal of Machine Learning Research, the list of references of 3 (2003) 583-617.
In another embodiment, be described for the sub-segmentation result that three kinds of features are corresponding, these three kinds of features are respectively color, texture and SIFT, and the sub-segmentation result based on these three kinds of features is expressed as the process of hypergraph as shown in Figure 4, Ke Yiwei:
With color mark vector C, texture markings vector T and SIFT label vector S, represent the sub-segmentation result utilizing color, texture and SIFT feature to obtain respectively.Such as C, the label vector C [3,1,2 ... ] T illustrates cluster classification belonging to every sub regions, as shown on the left of Fig. 4, subregion 1 belongs to the 3rd classification, and subregion 2 belongs to the 1st classification etc.If be identical based on the category label of two sub regions ri and rj under same feature, then show that this two sub regions has been merged into a region in the process obtaining sub-segmentation result.
For hypergraph G (V, E), the summit V of hypergraph just represents the subregion of pre-segmentation acquisition, r1, r2, and ri ∈ V.Set E represents super limit (hyperedges) and E={{C (p) }, { T (q) }, { S (r) } }.Each label vector has certain clusters number, and such as label vector C has P cluster, and each cluster is expressed as C (p), p=1, and 2 ..., P; Label vector T has q cluster, and each cluster is expressed as T(q), q=1,2 ..., q.The cluster classification of the label vector belonging to all subregion is expressed as matrix, as shown on the right side of Fig. 4, each column vector in this matrix, as C (p), T (q), and S (r), namely a super limit in hypergraph is represented, in each column vector 1 represents that corresponding row is that certain sub regions belongs to this cluster (super limit), and the subregion that the row of 0 expression correspondence is namely corresponding does not belong to this cluster (super limit), and the often row addition in matrix { C (p) } is all 1.Each sub-segmentation result can be expressed as the form of hypergraph by said process.
After each sub-segmentation result is expressed as hypergraph, clustering ensemble can be carried out based on this hypergraph antithetical phrase segmentation result, obtain the final segmentation result of image.The method of this clustering ensemble has multiple, and in one embodiment of this invention, adopt the method for spectral clustering to obtain final segmentation result, as shown in Figure 5, the method can comprise:
Step 501, sets final clusters number.
When carrying out spectral clustering and being integrated, need first to set clusters number, this clusters number can rule of thumb or other existing method arrange, clusters number setting etc. that also can be corresponding according to the sub-segmentation result of a certain feature.In the present embodiment, consider in segmentation, if adjacent similar subregion is all clustered into same object, similarity sum between region then in this object class should be minimum, and for a good segmentation result, namely all regions in image are all correctly clustered into several object, then the characteristic similarity sum of all clusters of this image is minimum.Because often kind of feature segmentation effect in son segmentation is different, optimum sub-segmentation result should be the sub-segmentation result that can produce minimum similarity degree sum, and this clusters number corresponding to sub-segmentation result will be used for the clusters number of final clustering ensemble.As shown in Figure 6, its detailed process can comprise:
Step 601, the sub-segmentation result corresponding at least two kinds of features calculates characteristic similarity respectively.
Example is characterized as with color, texture and SIFT tri-kinds, the characteristic similarity of the sub-segmentation result obtained under calculating each feature respectively, also the characteristic similarity between combined region rear region under a certain feature is namely calculated, this circular can with the step 301 in previous embodiment, 302 similar, repeat no more herein.
Step 602, is normalized calculating the characteristic similarity obtained.
Because every sub-segmentation result employs different features, therefore calculated characteristic similarity, can by Gaussian function being normalized the feature similarity calculated not in same dimension.
Step 603, using the clusters number of the sub-segmentation result corresponding to the minimum value of characteristic similarity after normalization as final clusters number.
After normalization, can using the clusters number corresponding to sub-segmentation result minimum for characteristic similarity as the clusters number of finally carrying out spectral clustering, such as, the characteristic similarity of sub-segmentation result corresponding under color characteristic is minimum, the then clusters number p of this sub-segmentation result, namely can be used as final number.
After determining the clusters number of spectral clustering, following steps 502 can be performed.
Step 502, carries out spectral clustering by each sub-segmentation result according to final clusters number based on hypergraph integrated, obtains the segmentation result of Given Graph picture.
Based on the principle of spectral clustering, when two regions are divided into a class, show that these two regions all merge under various features, therefore integrated result is that these two regions are merged into a region the most at last; On the contrary, merge in the sub-cutting procedure of a few features if there are two regions, and all do not merge under most of feature, then show, these two regions there are differences, and should not merge, then integrated result is that these two regions can not be merged.Such as, two green areas represent meadow and shrub respectively, and when based on color characteristic, merge, and all do not merge under texture and SIFT feature space, then integrated result does not also merge.Concrete carries out the process of spectral clustering similarly to the prior art according to the clusters number of hypergraph model and setting, can be called see name " On spectral clustering:analysis and an algorithm ", author is NG A Y, JORDAN M I, WEISS Y., is published in Advanced in Neural Information Processing System, 2002, the document of 2:849-856, repeats no more herein.
Final image segmentation result can be formed through said process.
The embodiment of the present invention by getting colors from global characteristics first respectively, textural characteristics, SIFT feature is selected from local scale invariant features, respectively based on color, texture, SIFT feature to Image Segmentation Using, then by a hypergraph model, the segmentation result under various features space is merged, defined one of image final division by the spectral clustering on hypergraph.This dividing method combines various features effectively to Image Segmentation Using; And can regard by being merged by the subregion of image the process being jumped to another node by a node according to the semantic similarity probability of hinged node as, not only consider the syntople in region, and the merging of neighboring region is converted to the merging process based on semanteme, greatly improve the performance of segmentation; By arranging the link weights of merging, adopting the strategy of greed, improving the speed of region merging technique; Further, by choosing the clusters number of spectral clustering, better segmentation effect is obtained.
Be more than the description to the inventive method embodiment, below the system realizing said method be introduced.
See Fig. 7, it is the structural representation of a kind of image segmentation system of the embodiment of the present invention.
This system can comprise:
Pre-segmentation unit 701, for carrying out pre-segmentation to Given Graph picture, obtains multiple subregion;
Merge cells 702, for merging described multiple subregion respectively according at least two kinds of features, obtains the sub-segmentation result that often kind of feature is corresponding;
Converting unit 703, for being expressed as hypergraph by sub-segmentation result corresponding at least two kinds of features;
Cutting unit 704, for the sub-segmentation result of at least two kinds of features being carried out clustering ensemble based on hypergraph, obtains the segmentation result of Given Graph picture.
First, pre-segmentation unit 701 pairs of Given Graph pictures carry out pre-segmentation, are partitioned into multiple subregion.The method of segmentation can adopt Ncut method Iamge Segmentation to be become several latticed join domains, then, merge cells 702 merges multiple subregions that pre-segmentation obtains respectively based on different features, obtain the sub-segmentation result that often kind of feature is corresponding, wherein, for the merging of subregion, webpage level algorithm can be used or other are similar to the mode of webpage level algorithm, as the random walk method based on similarity, subregion is merged.After merge cells 702 obtains multiple sub-segmentation result, sub-segmentation result corresponding for all features can be expressed as a hypergraph by converting unit 703.Wherein, all subregion that pre-segmentation obtains is expressed as the summit of hypergraph, sub-segmentation result corresponding for each feature is expressed as the super limit of hypergraph.Finally carry out clustering ensemble by cutting unit 704 according to this hypergraph, obtain the final segmentation result of this Given Graph picture.Wherein, the method that clustering ensemble adopts can be spectral clustering, also can be other clustering method.
In the embodiment of the present invention, the dividing method that this system realizes combines the segmentation of various features to image effectively, by a hypergraph model, the segmentation result under various features space is merged, and formed one of image final division by the clustering ensemble on hypergraph, make the performance of feature to segmentation serve due effect.
See Fig. 8, it is the structural representation of a kind of merge cells of the embodiment of the present invention.
In embodiments of the present invention, the merge cells of this image segmentation system can comprise:
Locator unit 801, for determining the link area of every sub regions;
Computation subunit 802, for calculating the merging weights of the link area of every sub regions;
Merge subelement 803, merge for the link area that every sub regions and its merging weights are satisfied condition.
Wherein, this locator unit 801 can further include:
Extraction module 8011, is adjacent the eigenwert of subregion for extracting subregion;
First computing module 8012, for the eigenwert according to extraction, calculating subregion is adjacent the characteristic similarity between subregion;
Determination module 8013, for characteristic similarity is greater than similarity threshold, and region area is greater than the link area of adjacent subregion as this subregion of subregion.
Computation subunit 802, specifically may be used for according to the characteristic similarity between every sub regions and its link area and semantic similarity valuation functions, according to webpage level algorithm, determines the merging weights of the link area of every sub regions.
Merge subelement 803, specifically may be used for the link area of every sub regions and its merging maximum weight to merge.
The merge cells of this system can regard by being merged by the subregion of image the process being jumped to another node by a node according to the semantic similarity probability of hinged node as, not only consider the syntople in region, and the merging of neighboring region is converted to the merging process based on semanteme, greatly improve the performance of segmentation; And by arranging the link weights of merging, adopting the strategy of greed, improving the speed of region merging technique.
See Fig. 9, it is the structural representation of a kind of cutting unit of the embodiment of the present invention.
In embodiments of the present invention, the cutting unit of this image segmentation system can comprise:
Setting subelement 901, for setting final clusters number;
Segmentation subelement 902, integrated for the sub-segmentation result of at least two kinds of features being carried out spectral clustering according to final clusters number based on hypergraph, obtain the segmentation result of described Given Graph picture.
Wherein, this setting subelement 901 can further include:
Second computing module 9011, for calculating the characteristic similarity of sub-segmentation result corresponding to often kind of feature;
Processing module 9012, the characteristic similarity for the sub-segmentation result at least two kinds of features is normalized;
Setting module 9013, for using the clusters number of the sub-segmentation result corresponding to the minimum value of characteristic similarity after normalization as final clusters number.
This cutting unit by choosing the clusters number of spectral clustering, and is defined one of image final division by the spectral clustering on hypergraph, obtains better segmentation effect.
The specific implementation process of each unit in above system please refer to the corresponding description of preceding method embodiment, repeats no more herein.
Above-described embodiment of the present invention, does not form limiting the scope of the present invention.Any amendment done within the spirit and principles in the present invention, equivalent replacement and improvement etc., all should be included within claims of the present invention.
Claims (11)
1. an image partition method, is characterized in that, comprising:
Pre-segmentation is carried out to Given Graph picture, obtains multiple subregion;
Respectively described multiple subregion is merged according at least two kinds of features, obtain the sub-segmentation result that often kind of feature is corresponding;
Sub-segmentation result corresponding for described at least two kinds of features is expressed as hypergraph;
Based on described hypergraph, the sub-segmentation result of described at least two kinds of features is carried out clustering ensemble, obtain the segmentation result of described Given Graph picture;
Described described multiple subregion to be merged, comprising:
Determine the link area of every sub regions;
Calculate the merging weights of the link area of described every sub regions, comprise: according to the characteristic similarity between described every sub regions and its link area and semantic similarity valuation functions, according to webpage level algorithm, determine the merging weights of the link area of described every sub regions;
The link area that described every sub regions and its merging weights satisfy condition is merged.
2. method according to claim 1, is characterized in that, the described link area determining every sub regions, comprising:
Extract the eigenwert that subregion is adjacent subregion;
According to the eigenwert extracted, calculate described subregion and be adjacent characteristic similarity between subregion;
Characteristic similarity is greater than similarity threshold, and region area is greater than the link area of adjacent subregion as described subregion of described subregion.
3. method according to claim 1, is characterized in that, described semantic similarity valuation functions is normal distyribution function or polygronal function.
4. method according to claim 1, is characterized in that, is describedly merged by the link area that described every sub regions and its merging weights satisfy condition, and comprising:
The link area of described every sub regions and its merging maximum weight is merged.
5. method as claimed in any of claims 1 to 4, is characterized in that, describedly based on described hypergraph, the sub-segmentation result of described at least two kinds of features is carried out clustering ensemble, obtains the segmentation result of described Given Graph picture, comprising:
Set final clusters number;
Based on described hypergraph, the sub-segmentation result of described at least two kinds of features is carried out spectral clustering according to described final clusters number integrated, obtain the segmentation result of described Given Graph picture.
6. method according to claim 5, is characterized in that, the final clusters number of described setting comprises:
The sub-segmentation result corresponding for described at least two kinds of features calculates characteristic similarity respectively;
Be normalized calculating the characteristic similarity obtained;
Using the clusters number of the sub-segmentation result corresponding to the minimum value of characteristic similarity after normalization as described final clusters number.
7. an image segmentation system, is characterized in that, comprising:
Pre-segmentation unit, for carrying out pre-segmentation to Given Graph picture, obtains multiple subregion;
Merge cells, for merging described multiple subregion respectively according at least two kinds of features, obtains the sub-segmentation result that often kind of feature is corresponding;
Converting unit, for being expressed as hypergraph by sub-segmentation result corresponding for described at least two kinds of features;
Cutting unit, for the sub-segmentation result of described at least two kinds of features being carried out clustering ensemble based on described hypergraph, obtains the segmentation result of described Given Graph picture;
Described merge cells comprises:
Locator unit, for determining the link area of every sub regions;
Computation subunit, for calculating the merging weights of the link area of described every sub regions, concrete, for according to the characteristic similarity between described every sub regions and its link area and semantic similarity valuation functions, according to webpage level algorithm, determine the merging weights of the link area of described every sub regions;
Merge subelement, merge for the link area that described every sub regions and its merging weights are satisfied condition.
8. system according to claim 7, is characterized in that, described locator unit comprises:
Extraction module, is adjacent the eigenwert of subregion for extracting subregion;
First computing module, for according to the eigenwert extracted, calculates described subregion and is adjacent characteristic similarity between subregion;
Determination module, for characteristic similarity is greater than similarity threshold, and region area is greater than the link area of adjacent subregion as described subregion of described subregion.
9. system according to claim 7, is characterized in that,
Described merging subelement, specifically for merging the link area of described every sub regions and its merging maximum weight.
10. according to the system in claim 7 to 9 described in any one, it is characterized in that, described cutting unit comprises:
Setting subelement, for setting final clusters number;
Segmentation subelement, integrated for the sub-segmentation result of described at least two kinds of features being carried out spectral clustering according to described final clusters number based on described hypergraph, obtain the segmentation result of described Given Graph picture.
11. systems according to claim 10, is characterized in that, described setting subelement comprises:
Second computing module, for calculating the characteristic similarity of sub-segmentation result corresponding to described often kind of feature;
Processing module, the characteristic similarity for the sub-segmentation result to described at least two kinds of features is normalized;
Setting module, for using the clusters number of the sub-segmentation result corresponding to the minimum value of characteristic similarity after normalization as described final clusters number.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201210224358.3A CN102819836B (en) | 2012-06-28 | 2012-06-28 | Method and system for image segmentation |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201210224358.3A CN102819836B (en) | 2012-06-28 | 2012-06-28 | Method and system for image segmentation |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN102819836A CN102819836A (en) | 2012-12-12 |
| CN102819836B true CN102819836B (en) | 2015-02-04 |
Family
ID=47303939
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201210224358.3A Expired - Fee Related CN102819836B (en) | 2012-06-28 | 2012-06-28 | Method and system for image segmentation |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN102819836B (en) |
Families Citing this family (15)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN103544697B (en) * | 2013-09-30 | 2016-01-20 | 南京信息工程大学 | A kind of image partition method based on hypergraph analysis of spectrum |
| CN103701647B (en) * | 2013-12-31 | 2016-08-31 | 武汉大学 | A kind of wireless network topology optimization generates method |
| CN103871089B (en) * | 2014-01-08 | 2017-01-18 | 天津大学 | Image superpixel meshing method based on fusion |
| CN105095921B (en) * | 2014-04-30 | 2019-04-30 | 西门子医疗保健诊断公司 | Method and apparatus for handling the block to be processed of sediment urinalysis image |
| CN104103044A (en) * | 2014-07-09 | 2014-10-15 | 上海电力学院 | Tackle cable jumping on-line detecting method based on K-Mean algorithm |
| CN104182981B (en) * | 2014-08-26 | 2017-02-22 | 北京邮电大学 | Image detection method and device |
| CN104820992B (en) * | 2015-05-19 | 2017-07-18 | 北京理工大学 | A kind of remote sensing images Semantic Similarity measure and device based on hypergraph model |
| CN104835174B (en) * | 2015-05-25 | 2017-12-15 | 厦门大学 | Robust Model approximating method based on Hypergraph model search |
| CN105279524A (en) * | 2015-11-04 | 2016-01-27 | 盐城工学院 | High-dimensional data clustering method based on unweighted hypergraph segmentation |
| CN105488800A (en) * | 2015-11-30 | 2016-04-13 | 上海联影医疗科技有限公司 | Feature extracting method and computer-aided diagnosis method and system |
| WO2017092615A1 (en) | 2015-11-30 | 2017-06-08 | 上海联影医疗科技有限公司 | Computer aided diagnosis system and method |
| CN105574532B (en) * | 2015-12-11 | 2019-03-08 | 深圳大学 | A kind of provincial characteristics based on binary segmentation tree describes method and system |
| CN112634290B (en) * | 2020-12-30 | 2023-09-19 | 广州南洋理工职业学院 | Graph segmentation method based on clustering interaction |
| CN113689431B (en) * | 2021-10-27 | 2022-02-11 | 常州微亿智造科技有限公司 | Industrial product appearance defect detection method and device |
| CN114838670A (en) * | 2022-03-11 | 2022-08-02 | 南京北新智能科技有限公司 | Photovoltaic panel dust detection method based on color analysis |
Citations (2)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN102360494A (en) * | 2011-10-18 | 2012-02-22 | 中国科学院自动化研究所 | Interactive image segmentation method for multiple foreground targets |
| CN102509097A (en) * | 2011-09-29 | 2012-06-20 | 北京新媒传信科技有限公司 | Method and device for image segmentation |
-
2012
- 2012-06-28 CN CN201210224358.3A patent/CN102819836B/en not_active Expired - Fee Related
Patent Citations (2)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN102509097A (en) * | 2011-09-29 | 2012-06-20 | 北京新媒传信科技有限公司 | Method and device for image segmentation |
| CN102360494A (en) * | 2011-10-18 | 2012-02-22 | 中国科学院自动化研究所 | Interactive image segmentation method for multiple foreground targets |
Non-Patent Citations (2)
| Title |
|---|
| A BAYESIAN MODEL FOR SCENE CLASSIFICATION WITH A VISUAL GRAMMAR;Xiaoru Wang;《Proceedings of IEEE IC-BNMT 2011》;20111030;310-314 * |
| Video Object Segmentation by Hypergraph Cut;Yuchi Huang;《IEEE Conference on CVPR 2009》;20090625;1738-1745 * |
Also Published As
| Publication number | Publication date |
|---|---|
| CN102819836A (en) | 2012-12-12 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN102819836B (en) | Method and system for image segmentation | |
| Csurka et al. | An efficient approach to semantic segmentation | |
| Tuggener et al. | Deep watershed detector for music object recognition | |
| Lim et al. | Sketch tokens: A learned mid-level representation for contour and object detection | |
| CN105389584B (en) | Streetscape semanteme marking method based on convolutional neural networks with semantic transfer conjunctive model | |
| Fergus et al. | Learning object categories from Google's image search | |
| Guillaumin et al. | Multimodal semi-supervised learning for image classification | |
| Lazebnik et al. | Affine-invariant local descriptors and neighborhood statistics for texture recognition | |
| Guillaumin et al. | Large-scale knowledge transfer for object localization in imagenet | |
| Larlus et al. | Category level object segmentation by combining bag-of-words models with dirichlet processes and random fields | |
| Zhu et al. | Could scene context be beneficial for scene text detection? | |
| CN101923653B (en) | Multilevel content description-based image classification method | |
| CN101877064B (en) | Image classification method and image classification device | |
| CN104346620A (en) | Inputted image pixel classification method and device, and image processing system | |
| CN103295032B (en) | Based on the image classification method of spatial Fisher vector | |
| CN102968637A (en) | Complicated background image and character division method | |
| Tighe et al. | Understanding scenes on many levels | |
| CN106504255A (en) | A kind of multi-Target Image joint dividing method based on multi-tag multi-instance learning | |
| CN104751175A (en) | Multi-label scene classification method of SAR (Synthetic Aperture Radar) image based on incremental support vector machine | |
| CN102902976A (en) | Image scene classification method based on target and space relationship characteristics | |
| CN110807485A (en) | Method for fusing two-classification semantic segmentation maps into multi-classification semantic map based on high-resolution remote sensing image | |
| CN104036021A (en) | Method for semantically annotating images on basis of hybrid generative and discriminative learning models | |
| CN107301426A (en) | A kind of multi-tag clustering method of shoe sole print image | |
| Liu et al. | Unsupervised object segmentation with a hybrid graph model (HGM) | |
| Gao et al. | A hierarchical image annotation method based on SVM and semi-supervised EM |
Legal Events
| Date | Code | Title | Description |
|---|---|---|---|
| C06 | Publication | ||
| PB01 | Publication | ||
| C10 | Entry into substantive examination | ||
| SE01 | Entry into force of request for substantive examination | ||
| C14 | Grant of patent or utility model | ||
| GR01 | Patent grant | ||
| CF01 | Termination of patent right due to non-payment of annual fee |
Granted publication date: 20150204 Termination date: 20200628 |
|
| CF01 | Termination of patent right due to non-payment of annual fee |